Run the Memdoor gateway on a remote server

Set up the VM#

On your computer, create a VM and connect. If you've already clicked Deploy or have a VM, skip the first command and use its name:

ssh exe.dev new --name=my-memdoor
ssh my-memdoor.exe.xyz

On the VM, install Memdoor, point it at exe.dev's LLM integration, and run the gateway as a service on port 8000:

curl -fsSL https://memdoor.ai/install.sh | bash
export PATH="$HOME/.local/bin:$PATH"
mkdir -p ~/.memdoor
cat > ~/.memdoor/env <<'EOF'
AI_GATEWAY_BASE_URL=https://llm.int.exe.xyz/v1
AI_GATEWAY_API=chat
AI_GATEWAY_TOKEN=exe-llm-integration
MEMDOOR_MODEL=accounts/fireworks/models/glm-5p3-flash
EOF
chmod 600 ~/.memdoor/env
memdoor gateway install --port 8000
git init ~/project
cd ~/project
MEMDOOR_GATEWAY=http://localhost:8000 memdoor run "Reply with exactly: OK" </dev/null

The installer puts memdoor in ~/.local/bin. memdoor gateway install writes a systemd user service, memdoor, that reads ~/.memdoor/env and starts again after a crash or reboot. The four lines in that file use Memdoor's company-gateway mode: every model call goes to exe.dev's integration, which adds the credentials itself, so the token is a placeholder. Memdoor then contacts nothing else: no OpenRouter, no memdoor.ai. AI_GATEWAY_API=chat matters: with Memdoor's default Responses API, follow-up turns stalled in our tests.

The test run creates Memdoor's first account on the gateway and saves its login in ~/.memdoor/local-login, which the ssh door reads from your computer. It should print OK. Check the gateway:

memdoor --gateway http://localhost:8000 gateway status

It says answering at http://localhost:8000 and service active.

Connect from your computer#

On your computer, install Memdoor (on Windows, irm https://memdoor.ai/install.ps1 | iex instead of the first line), ssh to the VM once to accept its host key, and save the gateway address:

curl -fsSL https://memdoor.ai/install.sh | bash
ssh exedev@my-memdoor.exe.xyz true
export MEMDOOR_GATEWAY='ssh://exedev@my-memdoor.exe.xyz?port=8000'

Put the export line in your shell's rc file to keep it. Memdoor runs your own ssh with BatchMode, so it can't answer a host-key question itself.

Give it a task with memdoor run. --dir names the directory on the VM:

memdoor run --dir /home/exedev/project "write a hello-world Go program with a test, and run the test"

The first command signs your computer in with the VM's login and stores it in ~/.memdoor on your computer. Then open the window:

memdoor --dir /home/exedev/project

The footer says live my-memdoor.exe.xyz:~/project. Files are read, commands run and tests pass on the VM. Close the window or your laptop and the gateway keeps the session; memdoor resume --last picks it up again. On a computer that has never signed in, open the window only after one memdoor run: in our tests the window alone answered Authentication required.

If you skip "Set up the VM", memdoor --gateway ssh://… installs Memdoor and the service on the VM by itself. But it then creates the first account from your computer and leaves the model up to you, so write the env file on the VM anyway to use exe.dev's models.

Watch from your phone#

In the window, type /remote. Memdoor prints a link and a QR code like http://my-memdoor.exe.xyz:8000/r/…#k=…. Change http to https and open it on a phone that's signed in to exe.dev: the conversation opens in a Memdoor terminal, with keys for Esc, Tab and the arrows, through the exe.dev HTTPS proxy. The gateway on the VM runs that terminal and is its own relay, so no Memdoor plan is needed and the link keeps working after you close the window on your computer. /remote off revokes it.

The proxy keeps port 8000 private to you. To let a teammate watch, share the VM with them and send the watch-only link Memdoor prints alongside.

Pick a model#

MEMDOOR_MODEL in ~/.memdoor/env is the model every turn uses. To switch, on the VM:

sed -i 's|^MEMDOOR_MODEL=.*|MEMDOOR_MODEL=accounts/fireworks/models/kimi-k3|' ~/.memdoor/env
systemctl --user restart memdoor

Other open-weight models on exe.dev include accounts/fireworks/models/glm-5p3, accounts/fireworks/models/minimax-m3 and accounts/fireworks/models/deepseek-v4p1-flash. /model in the window can't list them: Memdoor reads its model list from OpenRouter.

memdoor upgrade --yes